PDF file

COSMO data and conventional molecular descriptors
compared by redundancy analysis
Bettina Liebmann, Anton Friedl, Kurt Varmuza
Institute of Chemical Engineering, Vienna University of Technology, Getreidemarkt 9, Vienna, Austria.
bettina.liebmann@tuwien.ac.at, www.lcm.tuwien.ac.at, www.vt.tuwien.ac.at
Introduction
Correlation Between Variable Groups
Redundancy analysis (RA) allows comparing the multivariate information content
of variable blocks [1]. In quantitative structure-property relationships (QSPR),
these variable blocks are groups of molecular descriptors that originate from
diverse theoretical concepts. Whenever new types of descriptors are published, it
would be helpful to know if they cover similar aspects as already existing
descriptors (for a given set of structures). We apply RA to gain insight into the
correlations between the blocks of variables, and to assess the extent of
redundancy of three COSMO data groups [2] with 28 molecular descriptors
groups calculated by software Dragon [3,4]. The asymmetric redundancy index
allows not only to quantify the redundant information, but also to spot which of
the groups can model the other one. Selected groups are then used as X in a
QSPR task for the property y being the water-octanol partition coefficient, log P.
The correlations between the 31 groups have been quantified by the “summed
redundancy index RY” (see heatmap below) [1]. Potentially different information
of variable groups is indicated by a yellow square, which denotes very low
correlation. Group combinations with highest correlations are red. The COSMO
groups (G29-G31) have very high correlation (RY > 0.94) with “Burden eigenvalues” (G9), “RDF descriptors” (G16), and “WHIM descriptors” (G18). In addition,
the redundancy among the COSMO groups is high (RY > 0.92).
“Heatmap” of the summed redundancy indices
RY for 31 variable groups (G1 to G31), values
encoded by colour. Note the asymmetry!
yellow = no or very low correlation
red = high correlation
white = no RY values available (for X-groups
with m>n, here white columns for G7, 8, 17,
19).
COSMO groups are G29-31. Blue squares
indicate group combinations with the three
highest RY values.
RY value
What is COSMO-RS?
COSMO-RS (conductor-like screening model for real solvents) [2] is a quantum
chemistry based statistical thermodynamics model for the prediction of
thermodynamic properties of fluids and liquid mixtures.
Molecular interactions are calculated from the screening (polarisation) charge
densities (σ) on the molecular surface, which is defined as the response of an
electric conductor to the charge density of the molecule.
1,2-dichloroethane , C2H2Cl2
difluoroacetate , CHF2COO-
Two sample molecules with their
screening charge density σ mapped
on the molecular surface (“σsurface”).
blue … negative σ, strong positive polarity
red … positive σ, strong negative polarity
green/yellow … small σ, weak polarity
Each of the three Dragon groups have 90-124
variables, and an intrinsic dimensionality (80%
variance) of either 5 or 8. The COSMO groups
have 21-61 variables, and intrinsic dimensionalities of either 3 or 7.
Modelling COSMO Data From Descriptors?
Does a high RY imply that a good linear model can be derived for prediction of
one group from the other? For each COSMO group, three groups with highest RY
were selected. For each single COSMO variable used as y, a PLS1 model was
created with an entire variable group used as X. The prediction performance was
estimated by rdCV. Result: COSMO data are better modelled by another COSMO
group than any single conventional molecular descriptors group (from Dragon).
Model error vs. redundancy plot.
RY between COSMO σ-moments (G31)
used as Y and all other groups used as
X. Test set predicted errors from rdCV
for 22 single, autoscaled y-variables of
G31 and all other groups used as X in
PLS1.
Each molecule can be characterised by the probability distribution of the
screening charge densities σ, called σ-profile. The moments derived from the σprofile are called σ-moments, and comprise, e.g., total charge, electrostatic
interaction energy, acceptor and donor functions. However, several other
moments do not have a simple physical interpretation. The σ-potential can be
interpreted as the affinity of a solvent for the surface of polarity σ.
General trend:
For low RY , modelling Y from X group is
not possible, prediction errors are high.
For high RY , a good model performance
is possible, but not necessarily.
COSMO data for 1,2-dichloroethane:
•σ-profile with 61 variables (blue curve, G29)
•σ-potential with 61 variables (red curve, G30)
(61 equidistant σ-values from -0.03 to 0.03 e/A²)
•σ-moments with 22 variables (not shown, G31)
Data and Software
• n=131 chemical structures from organic molecules.
Data format: SMILES (from database [5]) and SDF with approximate 3D structures
including all H-atoms (SMILES → 2D SDF [6] → 3D SDF [7])
• m=2691 variables: COSMO data and molecular descriptors, 31 data blocks.
(G1-G28) Molecular descriptors calculated by software Dragon [4], mDragon = 2561
(1) Constitutional indices, (2) ring descriptors, (3) topological indices, (4) walk and path counts, (5) connectivity index, (6)
information indices, (7) 2D matrix-based descriptors, (8) 2D autocorrelations, (9) Burden eigenvalues, (10) P-VSA-like descriptors,
(11) ETA indices, (12) edge adjacency indices, (13) geometrical descriptors, (14) 3D matrix-based descriptors, (15) 3D
autocorrelations, (16) RDF descriptors, (17) 3D MoRSE descriptors, (18) WHIM descriptors, (19) GETAWAY descriptors, (20) Randic
molecular profiles, (21) functional group counts, (22) atom-centred fragments, (23) atom-type E-state indices, (24) CATS 2D, (25)
2D atom pairs, (26) 3D atom pairs, (27) molecular properties, (28) drug-like indices
(G29-G31) COSMO data extracted from software COSMOtherm [6], mCOSMO = 130
(29) COSMO σ-profile, (30) COSMO σ-potential, (31) COSMO σ-moments
• Calculations in R [8], using R-package “chemometrics” and R-code for
redundancy analysis cf. [1]. Repeated double cross validation, rdCV [9], used for
all PLS models.
• Property for QSPR model: water-octanol partition coefficient, log P, values from
-0.86 to 5.18 (mean 1.74, standard deviation 1.2) from online database [10].
QSPR Model for log P
The best QSPR models for log P are computed
from COSMO data, both single COSMO groups
and all three together. The good prediction
performance could not be reached with
Dragon groups; not even with the single
Dragon groups G9, 16, 18 that exhibit high RY!
Despite the redundant information content,
COSMO variables obviously contain additional
information for modelling the log P. Hence,
they can be considered valuable molecular
descriptors.
X groups
SEP
a
All groups (G1-31)
0.5
8
All Dragon groups (G1-28)
0.6
8
All COSMO groups (G29-31) 0.4
4
COSMO G29
0.4
8
COSMO G30
0.5
6
COSMO G31
0.4
4
Dragon G9
0.9
2
Dragon G16
1.1
1
Dragon G18
1.0
10
SEP, stand. deviation of prediction errors from test sets (rdCV)
a, optimum number of PLS components
References
[1] K. Varmuza, P. Filzmoser, B. Liebmann, M. Dehmer, Redundancy analysis for characterizing the correlation between groups of
variables - applied to molecular descriptors. Chemom. Intell. Lab. Syst. (2011), doi: 10.1016/j.chemolab.2011.05.013
[2] A. Klamt, COSMO-RS: From Quantum Chemistry to Fluid Phase Thermodynamics and Drug Design, Elsevier Science Ltd.,
Amsterdam, The Netherlands, 2005.
[3] R. Todeschini, V. Consonni, Handbook of Molecular Descriptors, Wiley-VCH, Weinheim, Germany, 2000.
[4] Dragon, Software for molecular descriptor calculation, version 6.0, by R. Todeschini et al., www.talete.mi.it
[5] COSMOthermX, version C2.1, release 01.10; COSMOlogic GmbH & Co. KG, Leverkusen, Germany, 2009.
[6] The Open Babel Package, version 2.3.0, http://openbabel.org
[7] Corina software, Molecular Networks GmbH Computerchemie, www.mol-net.de, Erlangen, Germany, 2004.
[8] R. A language and environment for statistical computing, version 2.12.2. R Development Core Team, Vienna, Austria, 2011.
www.r-project.org
[9] P. Filzmoser, B. Liebmann, K. Varmuza, Repeated Double Cross Validation, J. Chemom., 23, 160-171 (2009).
[10] Physical Properties Database PhysProp, www.srcinc.com/what-we-do/databaseforms.aspx?id=386
Acknowledgements
We thank Prof. Peter Filzmoser for contributing R code, and Dr. Andreas Klamt, COSMOlogic for providing COSMOthermX software.
Conferentia Chemometrica 2011 – Sümeg, Hungary, 18.-21.9.2011